pith:2QE6M6KH
Cornerstones or Stumbling Blocks? Deciphering the Rock Tokens in On-Policy Distillation
High-loss Rock Tokens in on-policy distillation resist training yet add almost nothing to reasoning performance.
arxiv:2605.09253 v2 · 2026-05-10 · cs.CL · cs.AI
Add to your LaTeX paper
\usepackage{pith}
\pithnumber{2QE6M6KHAOIPRHZWANQJSVLOZR}
Prints a linked badge after your title and injects PDF metadata. Compiles on arXiv. Learn more · Embed verified badge
Record completeness
Claims
strategically bypassing these ``stumbling blocks'' can significantly streamline the alignment process, challenging the necessity of uniform token weighting and offering a more efficient paradigm for large-scale model distillation.
The causal interventions used to measure negligible functional contribution to reasoning performance are valid and complete, and that high-loss tokens identified as Rock Tokens truly have no downstream effect on model outputs.
Persistent 'Rock Tokens' in on-policy distillation resist teacher corrections, consume large gradient norms, yet add negligible value to reasoning, allowing targeted bypassing to streamline alignment.
Formal links
Cited by
Receipt and verification
| First computed | 2026-06-02T01:03:48.286268Z |
|---|---|
| Builder | pith-number-builder-2026-05-17-v1 |
| Signature | Pith Ed25519
(pith-v1-2026-05) · public key |
| Schema | pith-number/v1.0 |
Canonical hash
d409e679470390f89f36036099556ecc5f6b72f6c63d0ddf0907eb3bcb265a29
Aliases
· · · · ·Agent API
Verify this Pith Number yourself
curl -sH 'Accept: application/ld+json' https://pith.science/pith/2QE6M6KHAOIPRHZWANQJSVLOZR \
| jq -c '.canonical_record' \
| python3 -c "import sys,json,hashlib; b=json.dumps(json.loads(sys.stdin.read()), sort_keys=True, separators=(',',':'), ensure_ascii=False).encode(); print(hashlib.sha256(b).hexdigest())"
# expect: d409e679470390f89f36036099556ecc5f6b72f6c63d0ddf0907eb3bcb265a29
Canonical record JSON
{
"metadata": {
"abstract_canon_sha256": "92339fc440ab4d2fbc71c519d06d3f52f4852462f92dd4ccd22dfb127e656fdb",
"cross_cats_sorted": [
"cs.AI"
],
"license": "http://creativecommons.org/licenses/by/4.0/",
"primary_cat": "cs.CL",
"submitted_at": "2026-05-10T01:41:43Z",
"title_canon_sha256": "f0fe8d86a87a7b9d7c45fa7b0042d6ff0e5c4180c3fceaf30d14b62f7deef40f"
},
"schema_version": "1.0",
"source": {
"id": "2605.09253",
"kind": "arxiv",
"version": 2
}
}